In 2026, as the global LLM race enters the era of "hundred-model warfare," a country with a population of only about 10 million has chosen a different path — not engaging in an arms race over parameter scale, but protecting the uniqueness and cultural sovereignty of a language. On July 1, the Portuguese government officially released AMALIA, the first open-source large language model developed specifically for European Portuguese.
I. Strategic Positioning: Not a "Portuguese ChatGPT," But a National Digital Infrastructure
The name AMALIA stands for "Automatic Multimodal Language Assistant with Artificial Intelligence." It pays tribute to the legendary Portuguese Fado singer Amália Rodrigues, deeply binding a national cultural symbol with cutting-edge technology.
At the launch event, Prime Minister Luís Montenegro explicitly stated: "The strategic autonomy of Europe, today more than ever, is closely linked to AI. This model will allow us to face the coming decades with greater sovereignty and less dependence."
Reform Minister Gonçalo Matias emphasized that Portugal cannot afford to fall behind in AI development. As the government continues to push for digital transformation of public services, ensuring technological sovereignty becomes increasingly critical.
Notably, the project lead made it clear: "This is not a ChatGPT." AMALIA will not be launched as a consumer-facing chat application. Instead, it serves as an open-source foundational technology platform for public institutions, enterprises, and researchers to build their own specialized AI applications.
This design reflects Portugal's strategic intent: embedding AI capabilities into the "capillaries" of national governance, rather than creating a product to compete head-on with Silicon Valley giants.
II. Technical Approach: Starting with 9B Parameters, 22B Version on the Horizon
AMALIA was developed over approximately 18 months by a team of over 60 researchers and students from five Portuguese universities and research institutions — a classic "national team" project.
Core Technical Parameters:
Initial Version: 9 billion (9B) parameters, built upon the European open-source foundation model EuroLLM-9B
Multimodal Capabilities: The current version already supports understanding and processing of text, images, and speech
Phase 2 Target: A 22 billion (22B) parameter version with new agentic capabilities to be launched later this year
Training Data: Pretrained on approximately 4 trillion Portuguese words, with around 5.8 billion tokens of high-quality European Portuguese data
Technical Highlight: The development team placed special emphasis on optimizing for "European Portuguese" rather than "Brazilian Portuguese." While the two are mutually intelligible, nearly all major global LLMs are trained predominantly on Brazilian Portuguese data. AMALIA addressed this by extracting large amounts of local data from Portugal's web archive, Arquivo.pt, and strictly filtering out ".br" domains (Brazilian domains) to build a training corpus focused on European Portuguese.
For evaluation, the team released a dedicated benchmark for European Portuguese. Experimental results show that AMALIA significantly outperforms general-purpose baseline models of similar scale on Portuguese-specific tasks.
III. Open Source and Funding: Transparent Ecosystem Under Apache 2.0
All code, training data, and model weights of AMALIA are open-sourced under the Apache 2.0 license, allowing commercial use. Any enterprise, citizen, or entity can freely download and use the model.
Financial Investment:
Initial Investment: €5.5 million (approximately $6.26 million), from the Portuguese Recovery and Resilience Plan (PRR)
Additional Investment: €1.5 million, bringing total investment to €7 million, with funding support extending through 2027 for model iteration and sovereign AI infrastructure development
The project is coordinated by NOVA University Lisbon, in partnership with Instituto Superior Técnico (Lisbon), the University of Coimbra, the University of Minho, and the University of Porto, with overall support from the Portuguese Foundation for Science and Technology.
In terms of computing infrastructure, AMALIA was trained leveraging European high-performance computing facilities, including the Deucalion supercomputer in Portugal and the MareNostrum 5 supercomputer in Barcelona, Spain.
IV. First Application Scenarios: From Museum Guides to Naval Decision Support
AMALIA has already been validated in real-world environments, with initial deployments focused on public sector and strategic domains:
Citizen Services: Digital government service assistant, to be integrated into the gov.pt mobile application
Culture: Virtual guides for Portuguese museums and monuments
Education: AI-powered teaching assistant to support lesson planning
Defense: Decision-support tool for the Portuguese Navy in critical operations
The government plans to first integrate AMALIA into the digital service channels of public administration, with gradual expansion to private sectors including banking, insurance, telecommunications, and industry.
V. Portugal's Position in the European AI Sovereignty Wave
The release of AMALIA is the latest milestone in Europe's "technological sovereignty" movement. Previously, European AI companies such as France's Mistral and Germany's Aleph Alpha had already carved out space in the US-dominated AI landscape.
AMALIA's uniqueness in the European AI landscape rests on three pillars:
1. Linguistic Specificity: Unlike general-purpose European models like Mistral, AMALIA is a model "tailor-made for a single language variant." Its value lies not in being "larger," but in being "more accurate" — in understanding the grammar, idioms, and cultural references of European Portuguese.
2. State-led Open-Source Model: Unlike commercial models reliant on venture capital, AMALIA is directly funded by EU recovery funds and the national budget, designed from the outset as public infrastructure rather than a commercial product.
3. Data Sovereignty in Practice: AMALIA's training data comes from Portugal's own national web archive, with processing and training conducted entirely within Europe, without dependence on US cloud services. This fully localized chain — from data to training to deployment — represents a complete implementation of data sovereignty.
The AMALIA project sends several noteworthy signals:
First, the "sovereignty race" in AI is shifting from "compute arms race" to "data and language sovereignty." As global LLM capabilities converge, the ability to accurately understand and generate language within a specific cultural context is becoming a new form of strategic asset. By choosing the niche of European Portuguese, Portugal has adopted a pragmatic "asymmetric competition" strategy.
Second, the "small country, big model" model has demonstrative value. With a population of only about 10 million and a mid-tier European GDP, Portugal has demonstrated that by focusing on linguistic uniqueness, integrating academic resources, and leveraging EU funding, it is possible to develop nationally relevant LLMs with practical applications. This offers a reference for other medium-sized countries — AI competition is not necessarily about "who has more parameters," but also about "who understands their own language and culture better."
Third, open source is a "must-have" rather than an "option" for sovereign AI. AMALIA's fully open-source nature is both a technical choice and a political one. When AI is embedded in citizen services and national defense, "auditability" and "controllability" matter more than "top performance." Open source allows the model to be publicly inspected, modified, and locally deployed — precisely the core requirements of sovereign AI.
The true test for AMALIA lies not in its launch ceremony, but in the next two years — whether it will evolve from a government "strategic project" into a "digital infrastructure" routinely relied upon by enterprises, universities, and government agencies. That will determine whether it becomes a milestone in Europe's AI sovereignty movement, or a beautifully crafted academic exhibit.